Person
1 entry · 31 May 2023
Rewarding each correct step of a solution, not just the final answer, produced a model that solved 78% of a representative subset of the MATH benchmark.
Ideas & essays